Strong Detector With Simple Tracker

Zongheng Tang, Yulu Gao, Zizheng Xun, Fengguang Peng, Yifan Sun, Si Liu, Bo Li; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 3047-3053

Unmanned aerial vehicle (UAV) tracking is a research direction with practical application value and has received sufficient attention in recent years. Challenges such as complex backgrounds, small targets, and motion blur in UAV tracking make it difficult to directly apply existing tracking or detection methods. For example, some state-of-the-art (SOTA) single-object tracking methods such as Ostrack perform poorly when encountering target disappearance or camera offset. Existing detection methods are also difficult to apply directly to this task. This paper proposes a detection-based method with cascading post-processing modules to solve this task. Our entire process includes generating detection candidate boxes, adjusting candidate box scores through video classification, connecting candidate boxes between different frames through a simple tracker, and determining moving targets in the video through background modeling, followed by single-object tracking as post-processing to adjust the results. We finally achieved first place in the 3rd Anti-UAV challenge track1 and top three in track2.